Rough-Fuzzy Hybrid Approach for Identification of Bio-markers and Classification on Alzheimer's Disease Data

A new approach is proposed in this paper for identification of biomarkers and classification on Alzheimer's disease data by employing a rough-fuzzy hybrid approach called ARFIS (a framework for Adaptive TS-type Rough-Fuzzy Inference Systems). In this approach, the entropy-based discretization technique is employed first on the training data to generate clusters for each attribute with respect to the output information. The rough set-based feature reduction method is then utilized to reduce the number of features in a decision table obtained using the cluster information. Another rough set-based approach is employed for the generation of decision rules. After the construction and the evaluation phases of the proposed rough-fuzzy hybrid system, the classification is carried out on the testing set of the given data. The experimental results showed that the proposed approach achieved compatible classification results on Alzheimer's disease data compared to results from other existing approaches in the literature. It can be concluded that the proposed rough-fuzzy hybrid approach is a novel approach in predictive data mining in clinical medicine in terms of utilizing 1) rough set-based approaches for feature reduction and rule generation, 2) a hybrid fuzzy system for pattern classification, and revealing 3) rules for prediction of diagnostic results.

I. Introduction

Research interest in predictive data mining for clinical medicine has been growing over the recent years. The objective of predictive data mining in this area is to build models from high-dimensional medical information and use these to predict diagnostic results on unseen medical data in order to support clinical decision-making. Approaches in predictive data mining may be applied to the construction of decision models for medical procedures such as prognosis, diagnosis, and treatment planning, which may be embedded within the clinical systems as systematic supporting components [13]. As listed in the recent literature [13], the following data mining approaches have been used in this area so far; decision trees (C4.5), conventional decision rules with values, logistic regression from statistics, artificial neural networks (ANN), support vector machines (SVM), the Naïve Bayesian classifier, Bayesian networks, the k-nearest neighbors.